Winner's-Curse-Aware Bid Scaling with Guaranteed Cost Targets

Automated ad bidders usually follow a simple rule: estimate what an impression is worth with a conversion-prediction model, multiply that estimate by one campaign-wide number, and bid the result. The multiplier is tuned during the day so that the campaign stays within budget and hits its cost-per-conversion target. The rule treats the model's estimate as if it were the true value. The auction then filters those estimates: the bidder wins exactly where its estimate is high relative to competitors. Wherever the model is unsure because it has seen little similar data, the impressions it wins are biased toward ones it has overestimated. This is the winner's curse. The bidder then pays for conversions that do not arrive, and the campaign misses its target.

This proposal treats a cost-constrained bidder as a selective predictor, meaning a model that chooses which inputs to act on and is judged only on those. From that view we take two concrete things. The first is a per-impression correction that shrinks bids in proportion to the model's own uncertainty about its estimate, not uniformly. The second is a way to set the campaign multiplier so that realized cost per conversion stays under target with a stated probability, not merely on average. The central prediction is that, at matched realized cost, this recovers conversions that a uniformly lowered multiplier throws away. The gain should appear exactly when the value model is data-starved and disappear when it is not.

Core Claim

The link between the two directions is structural. In a second-price auction, a bidder who bids $\alpha \hat v(x)$ wins impression $x$ whenever $\alpha \hat v(x) > c(x)$, where $\hat v$ is predicted conversion value and $c$ is the price set by the highest competing bid. Winning is therefore a selection function $g(x) = \mathbb{1}[\alpha \hat v(x) > c(x)]$, and $\alpha$ plays the role of a selection threshold. The average cost-per-action (CPA) constraint studied in Cost per action constrained auctions is then a ratio over the selected set:

$$\text{CPA}(\alpha) = \frac{\mathbb{E}[c(x)\, g(x)]}{\mathbb{E}[y(x)\, g(x)]} \le \text{CPA}^*,$$

Here $y$ is the realized conversion. This has the same form as the selective risk $\mathbb{E}[\ell\, g]/\mathbb{E}[g]$ in Selective Classification for Deep Neural Networks. Heymann shows that linear scaling is optimal when values are known. The selective-classification view asks the question that literature skips: what happens when the score being thresholded is itself an error-prone estimate?

The answer depends on the kind of uncertainty. Suppose $\hat v(x) = \mathbb{E}[y \mid x]$ exactly. Then conditioning on a win changes nothing, because the win depends only on $x$ and on competitor prices, which do not depend on the outcome once $x$ is known. The randomness of conversion itself, which we call aleatoric noise, therefore needs no correction in expectation. The bias comes from model error, which we call epistemic error: $\hat v(x) = \mathbb{E}[y\mid x] + \varepsilon(x)$, where $\varepsilon$ comes from finite training data. Selecting on large $\hat v$ selects positive $\varepsilon$. By the usual Gaussian selection argument, the bias among won impressions grows with the spread $\sigma_{\text{epi}}(x)$ of that error.

A campaign-wide multiplier can only cancel this bias on average. It over-shades impressions the model knows well and under-shades those it does not. We therefore propose the bid

$$b(x) = \alpha \,\big(\hat v(x) - \kappa\, \sigma_{\text{epi}}(x)\big)_+ ,$$

where $\sigma_{\text{epi}}$ is estimated from ensemble disagreement, $\kappa \ge 0$ is fitted on held-out data, and $(\cdot)_+$ clips at zero. That clipping means very uncertain impressions receive no bid; it is the only place hard abstention enters. The multiplier $\alpha$ is then chosen as the largest value whose upper confidence bound on realized CPA meets $\text{CPA}^*$ with probability $1-\delta$. This follows the threshold-inversion procedure of Geifman and El-Yaniv, adapted to a ratio of sums.

Relation to Prior Work

The auto-bidding line supplies the base mechanism and the main assumption we relax. Display Advertising with Real-Time Bidding (RTB) and Behavioural Targeting and Auto-bidding and Auctions in Online Advertising: A Survey derive the multiplier as an inverted budget shadow price. Bid Optimization by Multivariable Control in Display Advertising extends it to joint budget and cost-per-click constraints, with feedback controllers that track the dual variables. Lightweight Auto-bidding based on Traffic Prediction in Live Advertising replaces reactive control with traffic forecasting and an $\alpha v + \beta$ bid. All of these take $\hat v$ as exact. We keep their pacing machinery and change only what gets scaled and how the target is certified.

The paper that constrains the design most is Safe Online Bid Optimization with Return-On-Investment and Budget Constraints subject to Uncertainty. It proves that no online algorithm achieves both sublinear regret and sublinear constraint violations. Its lower-confidence-bound variant captured only about 56% of optimal revenue in synthetic runs. That result warns against blanket pessimism. We do not claim to escape the impossibility. Our claim is narrower: when uncertainty differs across impressions, spending conservatism only where the model is uncertain costs less than spending it everywhere.

Two papers motivate the selective side. Selective Classification via One-Sided Prediction shows that, when errors must be kept low, learning the selection score directly beats thresholding the plug-in confidence. At 0.5% target error on CIFAR-10 it reports 64.1% versus 59.3% coverage. This suggests that a learned correction may beat a hand-set $\kappa\sigma$ rule, and we test it as a variant. Trading via Selective Classification shows that abstention helps under trading frictions, but its abstention does not interact with a budget that has to be spent over the day.

The nearest auction-side uses of uncertainty are Conformal Online Auction Design and Selective Call Out and Real Time Bidding. The first uses conformal lower bounds to set seller reserve prices and explicitly leaves bidder budgets out of scope. The second has the exchange choosing which bidders to solicit. Both act for the seller, not the constrained buyer.

Research Program

The work has three parts.

The first part checks the premise. We need to show that, in realistic logs, the realized-minus-predicted gap on won impressions grows with epistemic uncertainty and not with aleatoric uncertainty. To do this we train value models on deliberately restricted slices of training logs, estimate $\sigma_{\text{epi}}$ from a five-member ensemble, and bin won impressions by $\sigma_{\text{epi}}$. We also estimate aleatoric spread with a separate head. If the gap is flat in $\sigma_{\text{epi}}$, the proposal fails early and cheaply.

The second part builds the certified multiplier. We use a finite-sample upper bound on the CPA ratio, obtained from separate concentration bounds on total cost and total conversions over a calibration window. We then invert that bound to pick $\alpha$, and let the existing pacing controller manage budget within the certified range. Traffic is non-stationary, with swings of up to 30x noted in BAT: Benchmark for Auto-bidding Task, so we calibrate on sliding within-day windows. The guarantee is stated only under exchangeability within a window. Its failure under drift is measured, not assumed away.

The third part examines the cheap-click tension. Yang et al. (2019) buy cheap low-value clicks to lower average CPC, while Yang et al. (2025) impose a lower CPC bound to avoid low-quality traffic. We test whether $\sigma_{\text{epi}}$ separates those two kinds of cheap inventory better than a price floor does.

Baselines

The naive reference is plug-in linear scaling. It bids $\alpha\hat v(x)$, with $\alpha$ set offline by solving the constraint on the calibration window under the assumption that $\hat v$ is exact and with no feedback correction. It shows how large the winner's-curse violation is.

The strongest competitor is a feedback controller on realized cost: M-PID from Bid Optimization by Multivariable Control in Display Advertising, and BiCB from Lightweight Auto-bidding based on Traffic Prediction in Live Advertising where the environment supports it. These are the strongest because they observe realized CPA and push $\alpha$ down until the target is met, so they already remove the average bias. The only thing left for us to win is the non-uniform allocation. The reported numbers are that M-PID reached a Value Ratio of 0.928 on the Taobao logs, against 0.892 for I-PID and 0.549 for standard feedback control. BiCB with perfect traffic prediction reached 99.93% of theoretical maximum revenue on the NeurIPS 2024 competition data. The BiCB figure assumes exact values, and our restricted-data setting removes that assumption. For BAT, no numerical result for M-PID, TA-PID or ALM is reported on Lacuna; only rankings are given.

Every method receives the same value-model ensemble, the same calibration windows, and the same grid budget for controller gains. We count it as a win if we achieve at least 3% more conversions at equal realized CPA, or equal conversions with fewer target-violating campaigns, across seeds.

Three ablations isolate what matters. Setting $\kappa=0$ inside the certified procedure separates the effect of certification from the effect of shading. Shading by aleatoric spread instead of epistemic spread tests the core mechanism. Replacing shading with hard abstention, bidding zero above a $\sigma_{\text{epi}}$ threshold, tests whether discrete gating adds anything over continuous shading. We also run an oracle that bids on true value, which fixes the ceiling.

Why not just…?

Why not just let the PID controller lower the multiplier until realized CPA is met? This is the real threat. On average it fixes the same bias we target. It must fail in one specific way: when $\sigma_{\text{epi}}$ varies strongly across impressions, a uniform cut gives up well-predicted inventory to pay for badly predicted inventory. The experiment settles this by comparing conversions at matched realized CPA as training data is progressively reduced. If PID matches us at every level of data scarcity, the per-impression correction is not worth its complexity and we will report that.

Why not just add a hard confidence gate in front of the bidder? A fixed gate on predicted conversion probability ignores pacing. The one-sided prediction results also show that plain thresholding of confidence is suboptimal when errors must be kept low. In a second-price auction, bidding zero is also just one point on the shading curve, so we expect continuous shading to weakly dominate. We include the hard gate as an ablation. If it matches shading on the first-price split, where lowering a bid also lowers payment, that is informative and not merely a null result.

Why not just recalibrate the conversion model, for example with isotonic regression? Global recalibration makes $\hat v$ correct on average over all traffic. The winner's curse lives in the subset the auction selects, and that subset is chosen by the error itself. We test this by applying recalibration before every baseline and checking whether the dependence of the bias on $\sigma_{\text{epi}}$ remains among won impressions.

Proposed Experiments

The main testbed is the NeurIPS 2024 auto-bidding data and environment described in Auto-Bidding in Large-Scale Auctions: Learning Decision-Making in Uncertain and Competitive Games. Its roughly 500 million records include predicted conversion probabilities. We treat the provided probabilities as ground truth and train our own value models on 1%, 10% and 100% of training days. This makes epistemic error controllable, and its true size is known.

We add the VCG split of BAT, which is second-price-like, with a cost-per-click constraint; its first-price split serves as a stress test outside the regime where linear scaling is exact. We also use AuctionNet trajectories from PlatformBid: An Auto-Bidding Benchmark from a Unified Advertising Platform's Perspective for the setting where many bidders compete at once. There we additionally report the paper's CPA-ratio variance, to see whether uncertainty-aware shading makes the market more or less stable.

The metrics are conversions at matched realized CPA, the fraction of campaigns violating target, how closely the empirical violation rate tracks the nominal $\delta$, and budget under-delivery.

The decisive checks are three. The gain over the controller baseline must shrink toward zero as training data grows to 100% and in the oracle setting. Shading by epistemic spread must beat shading by aleatoric spread. The recovered conversions must come from low-$\sigma_{\text{epi}}$ inventory that the uniformly lowered multiplier stopped winning. If all three hold, the result explains a specific, measurable failure of point-estimate bid scaling and shows how to fix it. If they fail, the negative result is still useful, because it would show that the reactive controllers the field already uses absorb value-model uncertainty.